Pāriet uz galveno saturu

About entity matching

Different data sources often use different naming standards to refer to the same object or entity. Cognite Data Fusion's (CDF) contextualization tools allow you to match entities originating from various source systems to the same entity in the CDF data model. The interactive tools combine machine learning, a powerful rules engine, and domain expertise to map and connect the information.

With the mapped information, you can build applications where users, for example, can click a component in a 3D model to see all the related time series data or ask for all the pressure readings along a flow line.

The entity matching model detects potential matches and calculates a match score by finding similarities between source and target fields. For instance, the model may suggest that the time series AA-123-PS-1B matches the asset name Y-123-PS-4D. A match is impossible if there are no similarities in the strings used for matching.

The entity matching model can only find matches when there are some similarities between the strings and will never return entities with no similarities as a suggested match, even with large amounts of training data.

You can match entities using CDF, the Cognite API, or the Cognite Python SDK.

TIP

To increase the speed and quality of the matching process, we recommend that you clean the source data before you start, for example, by removing universal substrings and unexpressive special characters.

Detecting matches

CDF uses string tokenization to break strings into substrings. We refer to these substrings as tokens. The tokens in the source and target entities are compared and used to calculate the similarity between THE two entities. By default, the model uses the name fields to find similarities, but you can configure the model to use any field in CDF.

Suppose you select the simple similarity scoring model to match entities. In that case, the model uses the regular expression \p{L}+|[0-9]+ to remove all punctuation or dash characters and split the string into tokens that consist only of letters or numbers. For instance, the string 11-PDN-26540J-60 splits into the tokens 11, PDN, 26540, J, and 60.

The entity matching model calculates similarity features by comparing strings based on the similarities between tokens and their order.

For medium to large data sets, it's impossible to compare all source and target combinations. To reduce the number of comparisons, the model uses a blocking step to remove target entities that are not similar to the source entity.

padoms

The model can only find matches if there are some similarities between the source and target strings. However, the entities may have hidden similarities, such as common abbreviations like PT for Pressure Transmitter. Replace the abbreviation with the spelled-out name or vice-versa to let the model find more matches.

Defining match scores

The entity matching model uses the featureType property to define the combination of similarity features created and used in the entity matching model.

Each feature produces one feature score between the source and target. A higher feature score means a more likely match but can't be interpreted as a probability.

The model calculates all features specified from featureType for each pair of fields in the matchField property. For instance, if featureType entails N individual features and matchField specifies M pairs of fields, the model calculates N x M features for each candidate match.

Be cautious about adding extra matchFields as they may impact performance if they contain little or no similarity information.

Calculating match scores

An unsupervised model or a supervised model calculates the match score using the N x M feature scores and uses the match score to decide whether a source and target entity is a match.

Unsupervised model

Use the unsupervised entity matching model if you don't have any or only a few known matches as training data. The unsupervised model is the default model and creates similarity feature scores between the sources and targets to return a weighted average as the match score.

The model uses a simple average, assigning the same weight to each feature. Therefore, it doesn't make sense to calculate many different features when using the unsupervised model, and matchFields should only include fields known to contain relevant similarities.

In addition, you should limit featureType to simple features such as Simple or Bigram.

Tip

An unsupervised model always outputs the same results when run on the same data. It's good practice to start with an unsupervised model for the first iterations to understand the data and get some verified matches before creating a supervised model.

Supervised model

Use the supervised entity matching model when you have some known matches that can be used to decide which features to weigh up or down to calculate the match score from the feature scores.

The supervised model often performs better with good training data since it can weigh the different features into fine-tuned criteria to determine a similarity between the source and target entities. It also allows you to use more matchFields by automatically determining which matchFields to weigh as stronger or weaker.

Tip

If you have only a small number of known matches or your matches are all based on the same pattern, the model might put a lot of weight on features that follow the pattern (overfitting). Overfitting may result in poor generalization, which means the model performs poorly on slightly different data that doesn't follow the pattern. The more varied your training data is, the better the model will pick up different patterns and perform better.

Jupyter Notebook tutorials